This would be sad if pre-training with feedback is actually better, because pre-training is by far the most expensive part of the training, and you wouldn't want to re-train from scratch every time you update your reward model.
RESEARCH
-
Pre-training scale reduces need for human feedback fine-tuning
By
–
Haven't read carefully, but Figure 1 suggests that with more pre-training, you need less fine-tuning with human feedback (which is usually the case with fine-tuning). So at scale this would make potentially no difference?
-
Fast AI Solutions: User Preferences and Ethical Implications
By
–
Sure. Still many folks prefer fastlane, since it’s much faster to get results. Do you blame them?
-
Sam Altman Debunks AI Rumor as Bullshit in Interview
By
–
Ese dato no sólo es una burrada, sino que además ha sido desmentido por el propio Sam Altman en la entrevista que pudisteis ver en mi canal. Literalmente calificó a este rumor como "bullshit"
-
AI Emotion and Consciousness: Smoke and Mirrors or Reality?
By
–
Hmm an Artificial Intelligence that can think and feel ?
Not just yet I think, but some good smoke and mirrors #ArtificialIntelligence #MachineLearning
#deeplearning #Microsoft https://
washingtonpost.com/technology/202
3/02/16/microsoft-bing-ai-chat-interview/
… -
3D-aware Conditional Image Synthesis from Carnegie Mellon
By
–
"3D-aware Conditional Image Synthesis" Authors: Kangle Deng, Gengshan Yang, Deva Ramanan, Jun-Yan Zhu from @CarnegieMellon Read more about their project here: https://
cs.cmu.edu/~pix2pix3D/ -
Learnable Prompts Emerge as Key GNN Pre-training Architecture
By
–
In conclusion, it looks like a neat architecture to pre-train GNNs. Learnable prompts are a handy addition and might re-appear in future architectures. Can't wait to become a graph prompt engineer.
-
Learnable Prompt Vectors Reduce Parameters for Downstream Tasks
By
–
It only uses the learnable prompt vector and not the GNN weights, which are frozen for downstream tasks. This reduces the number of parameters that need to be updated, improves the computational efficiency of task learning/inference, and reduces the reliance on labeled data.
-
Prompt Tuning Optimizes Downstream Task Efficiency and Accuracy
By
–
Prompt tuning involves optimizing the learnable prompt to improve the computational efficiency and accuracy of downstream tasks. It is based on the similarity of subgraphs and is formulated using prompt-assisted task-specific subgraph representations.
-
Learnable Prompts Enable Better Task-Specific Knowledge Extraction
By
–
Learnable prompts provide a better alternative to handcrafted prompts, and enable the extraction of the most relevant prior knowledge for each task. The prompts are a dimension-wise reweighting (or a linear transformation) of the node representations.